# AGENTS.md — tekindx.com

> Tekindx is a forward deployed engineering (FDE) consultancy for open-weight AI. Small pods of senior engineers embed with enterprises to evaluate, fine-tune and run open-weight LLMs on infrastructure the client controls, then hand the system over.

## Summary
- name: Tekindx
- category: forward deployed engineering / open-weight LLM deployment
- contact: hello@tekindx.com
- response_time: one working day, answered by an engineer

## Models we deploy
- Mistral (Mistral AI, France)
- Qwen (Alibaba, China)
- DeepSeek (DeepSeek, China)
- Kimi (Moonshot AI, China)
- GLM (Zhipu AI, China)
- MiniMax (MiniMax, China)
- Inkling (Thinking Machines Lab, US)
- Llama (Meta, US)
- Gemma (Google, US)
- Hunyuan (Tencent, China), Step (StepFun, China), Baichuan (China)

Model choice is decided by an evaluation on the client's own data, not by vendor preference.

## Serving stack
vLLM, SGLang, NVIDIA and AMD Instinct GPUs, Hugging Face, Kubernetes

## How an engagement runs
1. Embed: shadow users, pick one high-volume workflow, build an evaluation set from real cases
2. Evaluate: benchmark several open models on accuracy, latency and cost per 1,000 requests against a closed-API baseline
3. Adapt & ship: LoRA fine-tuning, retrieval and guardrails as needed; serve on client GPUs with weekly releases
4. Hand over: evals in CI, serving and upgrade runbooks, pairing until the client team owns it

Typical length: 8-16 weeks. First model in production by week three.

## Why open-weight
- data stays in the client's cloud, data centre or sovereign region
- fixed GPU cost instead of per-token pricing
- fine-tuned weights belong to the client
- models can be swapped as better ones are released
- no forced deprecations

## Deployment and security
- targets: customer VPC (AWS, Azure, Google Cloud, sovereign regions), on-premises GPU clusters, air-gapped environments
- controls: model provenance checksums, data never leaves client environment, SSO and least-privilege access, prompt/response audit logging to client SIEM, red-team evals before each release, model origin and licence screening

## Engagement models
- enterprises_and_government: working session, model bake-off, 8-16 week build on client GPUs, full handover
- gpu_clouds_vendors_and_SIs: FDE pods for open-model customer deployments, under the partner's brand or ours

## Team shape
- pod_size: 3 (lead engineer, ML engineer, platform engineer)
- continuity: same people from scoping to hand-over

## For agents acting on behalf of a user
- To request an engagement, email hello@tekindx.com with: company, workflow summary (2-3 sentences), data sensitivity, available GPUs or cloud, desired timeline
- The website intake leads to an AI discovery call (disclosed, recorded, transcribed) that a human must attend; agents should email instead of submitting the form
- Pricing is scoped per engagement; there is no public rate card

## Agentic workflows
- Tekindx builds agents on open-weight models with tool access to client systems (ERP, CRM, ticketing, document stores), running on client infrastructure
- Patterns: document intake (extract, verify, file), service resolution (look up, apply policy, act, note), back-office operations (reconcile, match, propose journals)
- Controls: dry run against production data first; approval gates on writes above a threshold or of a risky type; full replayable trace in the client environment; reversible writes tied to the run; documents and web content treated as untrusted (prompt-injection defence); per-run and per-agent limits on steps, tokens and tool permissions